A dynamic meta-model approach to genetic algorithm solution of a risk-based groundwater remediation design model

Approximation ("meta") models have been used in coupled water resources optimization and simulation models to improve computational efficiency. In most instances, multiple simulation runs have been done before the optimization, which are then used to fit an approximate model that is used for the optimization. In this study, we propose a dynamic meta-modeling approach, in which artificial neural networks (ANN) is embedded into a genetic algorithm (GA) optimization framework to replace time-consuming flow and contaminant transport models. Data produced from early generations of the GA are sampled to train the ANN. We propose a dynamic learning approach that periodically re-samples new solutions both to update the ANN and correct the GA's converging route. This allows the meta model to adapt to the area in which the GA is searching and provide more accuracy. The results show that a proper sampling strategy can benefit both GA's searching and ANN's retraining. In our test case, more than 90 percent of the numerical model calls were saved with no loss in accuracy of the optimal solution.